A dataset is a collection of structured examples or records used to analyze data or develop and evaluate machine-learning models. In AI projects, it is one of the central assets alongside labels, training runs, and deployments. section 7.4.1 section 7.6
AI Foundry projects keep datasets with labeling tasks, training jobs, and deployment definitions. This gives a team one place to manage data, control access, and track the assets behind a specific use case. section 7.4.1 section 7.4
A useful dataset is not merely large; it should also support responsible model development.
- Represent relevant groups and cases, rather than allowing one demographic, region, genre, or preference to dominate the training data. section 7.5.1
- Connect data to fairness checks during training, such as disparity metrics, reweighting, or threshold adjustments. section 7.5.1
- Record the dataset version and its relationship to feature transformations, model versions, and training runs so results remain traceable. section 7.5.2
Book recommendations
A recommendation project might collect reviews and reading histories across age brackets, regions, and reading preferences. In AI Foundry, the data can be connected from Blob Storage, supplied as JSONL for training, and monitored for fairness as reader behavior changes. section 7.5.1 section 7.6
